APter: Privacy Enhancement in Deep Learning Services following Principle of Least Privilege
Hao Wu · 2023
The data abuse concern has risen along with the widespread development of Deep Learning Services (DLS). Mobile users specifically worry about data privacy being compromised. Mitigating this new concern is demanding because it requires excellent balancing between data privacy protection and highly-usable service. Unfortunately, existing works do not meet this unique requirement. In this work, we propose the data privacy protection mechanism called APter. APter is a user-side DLS-input converter, and its outputs, although still good for inference, can hardly be reconstructed and labeled for new model training inferred privacy information. At the core of APter is a lightweight converter to minimize private information according to the DLS input. Moreover, adapting APter does not have to change the existing provider backend and DLS models. We conduct comprehensive experiments with our APter prototype on mobile devices and demonstrate that APter can substantially raise the bar of data abuse difficulty with little impact on the service quality and overhead.